This is happening to me as well on tensorflow-metal=1.2.0 installed via pip. I can confirm that the same model trains properly on a Windows CPU and Ubuntu with a CUDA GPU.